A 3D SEGMENTATION ALGORITHM FOR ELLIPSOIDAL SHAPES. Application to nuclei extraction.

A 3D SEGMENTATION ALGORITHM FOR ELLIPSOIDAL SHAPES.

Application to nuclei extraction.

Emmanuel Soubies¹, Pierre Weiss¹, Xavier Descombes²

1ITAV-UMS3039, Universit´e de Toulouse, CNRS, Toulouse, France.

2MORPHEME team, INRIA/I3S/iBB, Sophia-Antipolis, France.

esoubies@gmail.com, pierre.armand.weiss@gmail.com, xavier.descombes@inria.fr

Keywords: Nuclei segmentation, 2D and 3D images, graph-cuts, marked point processes, ellipses and ellipsoids, multiple object detection, multiple birth and cut, bio-imaging.

Abstract: We propose some improvements of the Multiple Birth and Cut algorithm (MBC) in order to extract nuclei in 2D and 3D images. This algorithm based on marked point processes was proposed recently in (Gamal Eldin et al., 2012). We introduce a newcontrast invariantenergy that is robust to degradations encountered in fluorescence microscopy (e.g. local radiometry attenuations). Another contribution of this paper is a fast algorithm to determine whether two ellipses (2D) or ellipsoids (3D) intersect. Finally, we propose a new heuristic that strongly improves the convergence rates. The algorithm alternates between two birth steps. The first one consists in generating objects uniformly at random and the second one consists in perturbing the current configuration locally. Performance of this modified birth step is evaluated and examples on various image types show the wide applicability of the method in the field of bio-imaging.

1 INTRODUCTION

Cell or nuclei segmentation in 2D and 3D is a ma- jor challenge in bio-medical imaging. New micro- scopes provide images at higher resolutions, deeper into biological tissues, leading to an increasing need for automatic cell delineation. This task may be easy in certain imaging modalities where images are well resolved and contrasted, but it remains mostly unre- solved in emerging fluorescent microscopes dedicated to live imaging such as confocal, bi-photon, or selec- tive plane illumination microscopes. These modali- ties suffer from multiple degradations such as light attenuation in the sample, heavy noise and spatially varying blur that make the segmentation task hard even for human experts

Our aim in this work is to propose a segmenta- tion algorithm robust to such situations. Since images are heavily deteriorated, standard methods aiming at finding contours based on a sole regularity assump- tion such as active contours or Mumford-Shah deriva- tives fail for the segmentation. This observation led us to introduce strong shape priors: cells are modelled as ellipses or ellipsoids that should fit the image con- tents. Unfortunately, adding geometrical constraints makes the optimization problems highly non convex and appeal for the development of new global opti-

Figure 1: Example of a SPIM image (Multicellular tumor spheroid).

mization methods.

Following recent works (Descombes et al., 2009;

Descombes, 2011; Gamal Eldin et al., 2012), we use randomized algorithms that allow to escape from lo- cal minima. These algorithms are based on marked point processes. The Marked Point Process (MPP) approach (Baddeley and Van Lieshout, 1993; Dong and Acton, 2007) consists in estimating a configura- tion of geometric objects (in our case ellipses or el- lipsoids) whose number, location and shape are un-

known. It has proved to be very efficient in numer- ous image analysis applications as it allows the com- bination of radiometric information with strong geo- metrics constraints on the objects but also at a global scale. Defined by a density against the Poisson pro- cess measure, its main advantage is to consider a ran- dom number of objects and can be considered as an extension of the Markov Random Field approach. A review of this approach and its applications can be found in (Descombes, 2011).

The objects are defined on a state spaceχ=I×M by their location and their marks (i.e. geometric at- tributes). The associated marked point processX is a random variable whose realisations are random con- figurations of objects. Considering a Gibbs process, the modeling consists of an energy construction. Sim- ilarly to the Bayesian framework, this energy can be written as the sum of a data term and a prior. In this paper we consider a pairwise interactions prior that forbids intersections between objects. Once the model defined, the solution is obtained by minimiz- ing the energy. This energy being highly non-convex requires stochastic dynamics, such as MCMC meth- ods, to be minimized. The Reversible Jump MCMC embeded in a simulated annealing framework is a natural candidate for this task (Green, 1995). How- ever, in case of simple constraints such as non over- lap, the recently proposed multiple birth and death al- gorithm is preferable (Descombes et al., 2009). To avoid the fastidious calibration of annealing parame- ters, we propose to revisit the combination of the mul- tipe births principle with the graph cut paradigm pro- posed by (Gamal Eldin et al., 2012).

The paper is organized as follows. We formalize the segmentation problem as a minimization problem in section 2. Section 3 begins by a global algorithm description and is followed by a precise description of each algorithm step. We finish by presenting numeri- cal results in section 4.

2 PROBLEM STATEMENT

Figure [1] contains typical examples of images en- countered in biology. It is readily seen from these images that most nuclei contours can be well approx- imated by ellipses or ellipsoids, at least at a coarse scale. Moreover these nuclei cannot overlap due to obvious physical considerations. We thus formulate our segmentation problem as that of finding a set of non overlapping ellipsoids that fit the image contents.

We formalize this statement in the latter.

LetCn,n∈Ndenote the set of configurations con- tainingn objects that do not overlap. An element

x∈Cn is a set ofn non overlapping objects. Since the number of nuclei in the configuration is unknown, we aim both at finding this number n^∗ and the best configurationx∈Cn^∗ with respect to a certain data fi- delity term f(x). Our optimization problem can thus be formulated as follows. Let

g(n) =min

x∈Cn

f(x)

denote the minimum value of fin the setCn. We wish to find both

n^∗=arg min

n∈N

g(n) and

x^∗=arg min

x∈Cn∗

f(x).

By convention, we assume that C0 = 0/ and that

x∈minC0

f(x) =0. The data term f should thus be neg- ative for configurations that are likely to represent the nuclei parameters and positive otherwise. We detail how the ellipses are parametrized and the construc- tion of such a function in the following paragraphs.

Object modelling In 2 dimensions, ellipses are pa- rameterized using 5 parameters (see Figure 2):

• (x,y)∈Ω: center coordinates which should be- long to the image domainΩ.

• θ∈[0,2π[: angle with the horizontal direction.

• 0<λ−<b<a<λ+: describe the ellipses minor and major axes size. λ−andλ+are user defined parameters that describe the nuclei maximal size and ellipticy.

Figure 2: Parameters of the ellipse.

In 3 dimensions, nuclei are parameterized using 9 parameters:

• (x,y,z)∈Ω: center coordinates.

• φ,θ,ψ∈[0,2π[³: Euler angles to define the ellip- soids orientations.

• 0<λ−<c<b<a<λ+: axes lengths.

Overall, it can be seen that objects belong to a state spaceχdefined as a parallelepiped:

χ=Ω×[0,2π[×[λ−,λ+]² (1) in 2D and

χ=Ω×[0,2π[³×[λ−,λ+]³ (2) in 3D.

In this paper, the objects are denotedωand their boundary is denoted∂ω.

Data term Letu:Ω→Rdenote a grayscale image.

In order to define the data term f(x), we associate an elementary energyU_d(ω,u)to each elementω∈xand set:

f(x) =

∑

ω∈x

U_d(ω,u). (3) The functionU_d(ω,u)∈[−1,1]should be negative if the objectωis well positioned on the image and pos- itive otherwise.

In fluorescence microscopy, nuclei are usually characterized by bright region surrounded by a dark background since they are stained or genetically mod- ified in order to express a fluorescent protein. Unfor- tunately, their radiometry is not constant due to local bleaching or light attenuation in the deepest layers.

We thus need to construct an energy that iscontrast invariant, meaning that local modifications of the ra- diometry shall not affect the energy. Such an energy can be constructed easily by considering the normal to the image level lines_|∇u|^∇u where∇udenotes the usual gradient inR^d and|∇u|denotes the gradient magni- tude in the standard Euclidean norm. This tool is well known to be contrast invariant. Let us define an en- ergyUfor a given objectωas:

U(ω) = 1

|∂ω|

Z

∂ω

h ∇u(x) p|∇u(x)|²+ε²

,n(x)idx (4) whereh·,·idenotes the standard scalar product,|∂ω|

denotes the length of the object boundary, n(x)de- notes the outward normal toωat locationx∈∂ωand εis a regularization parameter that discard faint tran- sitions. The behavior of this energy is illustrated on Fig. 3. Overall, it does what is expected, but as can be seen on the illustration b) and d) in Fig. 3, badly located ellipses might have a negative energy and be kept in the final configuration. It is thus necessary to modifyUin order to promote well located objects only. A simple way to do so consists in setting:

U_d(ω,u) =ψ(U(ω),s)

wheres∈]−1,0]is an acceptance threshold for the objects. and

ψ(α,s) =min( 1

s+1α− s s+1,1).

Figure 3

Figure 4: Graph of the functionψ(α,s)with respect toα fors=−0.5. Note that the function becomes positive for values ofα>s.

This function is illustrated on Figure 4

Other data terms based on the contrast between the object interior and the background as presented by (Gamal Eldin et al., 2012) (in dimension 2) could also be used but present two drawbacks: first they re- quire to compute an integral over the interior of the domain while the proposed approach consist in com- puting a boundary integral which is faster. Second, such measures might be inaccurate in the case of very dense media, where the background can be difficult to extract. Finally our measure is contrast invariant, which is central for the targeted applications.

3 MULTIPLE BIRTH AND CUT ALGORITHM (MBC)

The Multiple Birth and Cut algorithm (MBC) has been proposed by (Gamal Eldin et al., 2012) for counting flamingos in a colony. In this section, we describe the different steps of the MBC algorithm (Al- gorithm 1).

The main idea consists in generating two random configurations of non-overlapping objects x and x⁰ (birth step) and then keep the subset of objects in x∪x⁰that minimizes f (cut step). This process is it- erated and decreases f at each iteration. The cut step can be performed efficiently using a Graph Cut algo- rithm (Boykov et al., 2001; Kolmogorov and Zabih, 2004). We describe this algorithm more formally be- low:

Interestingly, this algorithm contains only one pa- rameterN(the number of objects generated in a con-

Algorithm 1Multiple Birth & Cut algorithm Require: N

1: Generate a configurationx_[0]with Algorithm 2 2: n←−0

3: while(Not converged)do

4: Generate of a new configurationx⁰ using Algorithm 2.

5: x_[n+1]←−Cut(x_n∪x⁰) 6: n←−n+1

7: end while

figuration). We observed experimentally that this pa- rameter might affect slightly the speed of convergence but not the segmentation accuracy. This algorithm is thus much easier to tune than more standard RJM- CMC based dynamics.

3.1 Birth step

A new configuration x⁰ of non-overlapping objects is generated. Note that only objects which are in the same configuration have to respect the non- overlapping constraint, but two objects in different configurations can intersect as can be seen on Figure 5.

Figure 5: Two configuration on an image (the black ellipses are the object to detect)

The birth step is detailed in Algorithm 2. The fourth step of this algorithm can be efficiently imple- mented using a lookup table and the fast intersection algorithm proposed in the latter.

3.2 Cut step

This step consists in selecting the best configuration of non-overlapping objects in(x_[n]∪x⁰). To perform this optimization, a weighted graph is constructed.

The nodes of this graph are the objectsωof the two configurationsx_[n]andx⁰. This graph also possesses

Algorithm 2Birth step Require: N,nmax.

1: Setk=0,n=0,x⁰=0./ 2: whilek<Nandn<n_maxdo

3: Construct an objectω⁰by generating a random vector uniformly inχ.

4: Ifω⁰intersects an object inx⁰, setn=n+1 and go back to 3.

5: Otherwise setx⁰=x⁰∪ {ω⁰},k=k+1,n=0 and go back to 3.

6: end while

two special nodes, the source ’s’ and the sink ’t’. The weights should belong to [0,1]∪ {+∞} and are de- fined using the data termU_d(ω,u)by:

W(ω) = (1−U_d(ω,u))/2.

Graph construction

Each object of the configuration (x_[n]∪x⁰)is linked to the source and the sink. The difference between the objectsω_i∈x_[n]and the objectsω_j∈x⁰is that the objectsωi∈x_[n]are linked to the source with a weight equal to the data energyW(ω)and to the sink with a weight equal to 1−W(ω), while it is the reverse for the objectsω_j∈x⁰.

The weights associated to edges linking two ob- jects are non zero only when two objects intersect.

If ω1 ∈x_[n] (current configuration) intersects with ω2∈x⁰ (new configuration), the link fromω1toω2

is set to∞and the link fromω2toω1is set to zero¹. This ensures that the cut step generates an admissible configuration (with no overlapping objects). Figure 6 summarises the graph construction of the configura- tions on Figure 5. The nuclei to detect are represented by black ellipses.

Cut

Once the graph is constructed, we perform a cut that consists in partitioning the vertices into two disjoint subsets. One contains the source and the other the sink. The cut realized is the one with minimal cost (the one minimizing the sum of the weights of the re- moved edges).

After the cut step, ifωi∈x_[n]is in the sub-graph containing the source, we keep it, otherwise we re- move it. On the contrary the objects ω_j ∈x⁰ are

1When two objects intersect the link affected by a weight of∞is always the link from the object of the cur- rent configuration to the object of the new configuration.

Figure 6: Graph corresponding to the figure 5 kept only if they belong to the sub-graph that con- tains the sink. This difference of interpretation be- tween the two configurations combined with the dif- ferent weights to the source and the sink, ensure that in case where an object ofx_[n]and an object ofx⁰in- tersect, only one can be kept.

The cut step is implemented using the graph-cut code developed by Yuri Boykov and Vladimir Kol- mogorov in (Boykov et al., 2001; Boykov and Kol- mogorov, 2004; Kolmogorov and Zabih, 2004).

3.3 A FAST DETERMINATION OF ELLIPSES INTERSECTION

One of the proposed algorithm bottleneck is the fast determination of whether two ellipsoids intersect or not. In this section, we present a fast algorithm to answer that question and prove theoretically that only a few arithmetic operations suffice to provide the answer with a low error rate.

Let ω be an ellipse or an ellipsoid. It can be defined using a quadratic functionQ_ω as ω={x∈ R^d,Q_ω(x) =1}. The quadratic functionQcan be de- fined by:

Q_ω(x) =hA(x−c),(x−c)i (5) wherecdenotes the object center and A is positive definite matrix defined by:

A=P⁻¹DP=P^TDP.

wherePis a rotation matrix andDis a positive diag- onal matrix. In 2D,Pis defined by:

P=

cos(θ) sin(θ)

−sin(θ) cos(θ)

and

D= ₁

a² 0

0 ¹

b²

.

In 3D, the notation become cumbersome and we leave them to the reader.

Letω1 andω2 be two ellipses or ellipsoids. In order to know whether they intersect or not, we can find the minimal level set ofQ_ω2 which intersects the boundary of ω1. If this level set is associated to a value greater than 1, the ellipses are separated, other- wise they overlap. This idea can be formulated as the following minimization problem:

min

x∈R^d,Q_ω1(x)≤1Qω₂(x) (6) This problem consists of minimizing a quadratic func- tion over convex set. Projected descent methods can thus be used. Unfortunately, there exists no closed form solution to the problem of projection of a point on an ellipse. We thus need to simplify the constraint set:

min

Qω1(x)≤1Q_ω2(x)

= min

hA₁(x−c₁),(x−c₁)i≤1hA₂(x−c₂),(x−c₂)i

= min

h√

A₁(x−c₁),√

A₁(x−c₁)i≤1hA₂(x−c₂),(x−c₂)i min

ky−√ A₁c₁k²₂≤1

hA₂(A⁻

1 2

1 y−c2),(A⁻

1 2

1 y−c₂)i.

wherey=√

A₁x. In this reformulation, the constraint setY={y∈R^d,ky−√

A1c1k²₂≤1}is a simplel²-ball and the function F(y) =hA₂(A⁻

1 2

1 y−c₂),(A⁻

1 2

1 y− c₂)iis a strongly convex differentiable function. We can thus use a projected gradient descent that writes:

Algorithm 3Detection of overlapping ellipsoids Require: Q_ω1,Q_ω2,ε>0.

1: Setk=0,y₀=^c1+c₂ ². 2: Setµ=^b21

a²₂,L=^a21

b²₂. 3: Setτ=_µ+L² .

4: while ky_k+1−y_kk ≥εdo 5: y_k+1

2 =yk−τ∇F(y_k).

6: y_k+1=ΠY

y_k+1

2

. 7: k=k+1.

8: end while

9: IfF(y_k)>=1 return 0 (the ellipsoids do not in-

tersect with high probability).

10: IfF(y_k)<1 return 1 (the ellipsoids intersect).

Let y^∗ denote the solution of the above prob- lem. The previous algorithm comes with the follow- ing guarantees:

Theorem 1. After k iterations, y_ksatisfies:

F(y_k)−F(y^∗)≤µ

2ky₀−y^∗k²₂

Q_F−1 Q_F+1

2k

ky_k−y^∗k²₂≤ ky₀−y^∗k²₂

Q_F−1 Q_F+1

2k

where

Q_F=a²₁ b²₂

a²₂ b²₁ ≤λ⁴₊

λ⁴₋.

Proof The Hessian of F isH_F(y) =2A⁻

1 2

1 A₂A⁻

1 2

1 . SinceA₁andA₂are products of orthogonal and diag- onal matrices (A=P^TDP), the eigenvalues ofHF(y) can be easily bounded:

λmin[H_F(y)]≥b²₁

a²₂ λmax[H_F(y)]≤a²₁ b²₂ The functionFis thusµ-strongly convex withµ≥b²₁

a²₂ and its gradient is L-Lipschitz with L≤ a²₁

b²₂. Us- ing standard convergence theorems in convex analysis (Bertsekas, 1999), we obtain the announced result.

The conditioning numberQF depends solely on the ratio between the major axis and the minor axis sizes and not on the dimensiond. This algorithm will thus be as efficient in 3D as in 2D. For two circles the ratioQ_F is equal to ^a_b=1 and the algorithm pro- vides the exact answer after one iteration. For elliptic ratios of 2,Q_f =16 and in the worst case, after 18 iterations, the algorithm returns a pointy^kthat is 100 times closer to the intersection thany₀. We also tested an accelerated algorithm by (Nesterov, 2004), where the convergence rate is of order^√

QF−1

√Q_F+1

2k

but it did not improve the computing times.

In our problems the ratio betweena andb is al- ways less than 2 and the algorithm usually converges in just a few iterations (2 to 10 depending on the prob- lem).

3.4 Acceleration by local perturbations

When the objects variability is important, the state space size increases and affects the convergence speed of the MBC algorithm. This problem is particulary important in 3D since ellipsoids are defined by 9 pa- rameters instead of 5 for the 2 dimensional case.

In order to improve the convergence speed, (Gamal-Eldin et al., 2011) proposed to insert aselec- tion phasein the birth step. This selection phase con- sists in generating a dense configuration of objects at

similar locations and to keep the best ones using Be- lief Propagation in order to form the new configura- tion.

In this paper, we propose another heuristic in or- der to increase the convergence speed. We propose to alternate between two different kinds of birth steps.

The first one is that proposed in algorithm 2. The sec- ond one consists in perturbating locally the current configuration. This principle mimics the proposition kernels used in RJMCMC algorithms (Perrin et al., 2005). The idea behind this modification is that after a while, most objects are close to their real location and that local perturbations may allow much faster convergence than fully randomized generation. This algorithm is described in details in Algorithm 4.

Algorithm 4MBC algorithm with local perturbations Require: N

1: Generate a configurationx_[0]using Algorithm 2.

2: n←−0

3: while(Not converged)do

4: Generate a uniformly distributed random num- berr∈[0,1].

5: ifr<p then

6: Generate a new configurationx⁰ using Algorithm 2.

7: else

8: Generate a new configurationx⁰ using Algorithm 5.

9: end if

10: x_[n+1]←−Cut(x_n∪x⁰) 11: n←−n+1

12: end while

Algorithm 5Birth step with local perturbation Require: x_[n].

1: whilek<size(x_[n])do

2: Construct an objectω⁰by local perturbation of ω_k∈x[n].

3: Ifω⁰intersects an object inx⁰, setk=k+1 and go back to 2.

4: Otherwise setx⁰=x⁰∪ {ω⁰},k=k+1 and go back to 2.

5: end while

Local perturbations

A given objectω inx_[n] is described by a set of pa- rametersλ∈χ(see equations 1 and 2). We generate the new objectω⁰by setting its parametersλ⁰=λ+z wherezis the realization of a random vector Z dis-

tributed uniformly inχ_εwhere :

χε= [−δ_xy,δxy]²×[−δ_ab,δ_ab]²×[0,2π[

in 2D and

χ_ε= [−δ_xyz,δ_xyz]³×[−δ_abc,δ_abc]³×[0,2π[³ in 3D.

The value of the differentδdescribes the perturba- tion extent. We observed that small values accelerates the convergence speed.

Comparison of the convergence speed

We have tested this method in order to compare the speed of convergence of the MBC algorithm and the MBC algorithm with local perturbation. Figure 7 presents the energy evolution with respect to time for both MBC and MBC with local perturbations (de- noted MBC with LP) on the same image (the 3D nu- clei of Drosophila embryo). The segmentation result is presented on Figure 14 (image size 700×350× 100). These results show that the MBC with LP algo- rithm strongly improve the MBC algorithm.

Figure 7: Comparison of the MBC and MBC with LP al- gorithms

4 RESULTS

In this section, we present some practical results in 2D and 3D. Figure 9 shows the segmentation result on a Drosophila embryo obtained using SPIM imaging.

This is a rather easy case, since nuclei shapes vary little. The images are impaired by various defects:

blur, stripes and attenuation. Despite this relatively poor image quality, the segmentation results are al- most perfect. The computing time is 5 minutes using a c++ implementation. The image size is 700×350.

Figure 8: 2D nuclei of Drosophila embryo

Figure 9: 2D segmentation of a nuclei of Drosophila em- bryo (Fig 8)

Figure 10 presents a more difficult case, where the image is highly deteriorated. Nuclei cannot be iden- tified in the image center. Moreover, nuclei variabil- ity is important meaning that the state space sizeχis large. Some nuclei are in mitosis (see e.g. top-left).

In spite of these difficulties, the MBC algorithm pro- vides acceptable results. They would allow to make statistics on the cell location and orientation, which is a major problem in biology. The computing times for this example is 30 minutes.

Figure 10: 2D segmentation of a multicellular tumor spheroid (Fig 1)

Nuclei segmentation is a major open problem with a large number of other applications. In Figure 11, we attempt to detect the spermatozoon heads. The foreseen application is tracking in order to understand

their collective motion. Figure 12 presents a multi- cellular spheroid surrounded by a circle of pillars. A typical problem consists in determining whether the pillars have moved afetr a while, due to the tumor dy- namic. They thus need to detect the pillars and eval- uate their motion. Figure 11 and Figure 12 show that the segmentation results are extremely precise.

Figure 11: Segmentation of a spermatozoon colony (5 min- utes). Image size: 2000 x 1024.

Figure 12: Nano pillars detection (less than 1 minute). Im- age size: 840 x 800.

3D results are presented in Figure 14 and 16. For the Drosophila embryo, the segmentation is very close to what a human expert would do. The computing times are 2 hours and the image size is 700×350× 100. The curves in Figure 14 correspond to this im- age.

The spheroid segmentation presented in Figure 16 is less precise than the previous ones due to an im- portant cell variability and to the fact that the images are extremly blurry in the Z direction. For that case, image restoration algorithms and the design of new energies robust to strong perturbations seem impor- tant.

Figure 13: 3D Drosophila embryo nuclei

Figure 14: 3D segmentation of the Drosophila embryo nu- clei (Fig 13)

Figure 15: 3D multicellular tumor spheroid

ACKNOWLEDGEMENTS

This work was partially funded by the Mission pour l’interdisciplinarit´e from CNRS, R´egion Midi Pyr´en´ees, PEPII CASPA3D and ANR SPHIM3D.

The authors wish to thank F. Malgouyres and J. Fehrenbach for interesting discussions. They also thank Val´erie Lobjois, Charlotte Emery, Jacques

Figure 16: 3D segmentation of the multicellular tumor spheroid (Fig 15)

Thomazeau, Paul Escande and Bernard Ducommun for their help in this project. They thank all the ITAV staff for their warm welcome in a biology laboratory.

REFERENCES

Baddeley, A. and Van Lieshout, M. (1993). Stochastic ge- ometry models in high-level vision. Journal of Ap- plied Statistics, 20(5-6):231–256.

Bertsekas, D. (1999). Nonlinear programming.

Boykov, Y. and Kolmogorov, V. (2004). An experimental comparison of min-cut/max-flow algorithms for en- ergy minimization in vision. IEEE Trans. Pattern Anal. Mach. Intell., 26(9):1124–1137.

Boykov, Y., Veksler, O., and Zabih, R. (2001). Fast ap- proximate energy minimization via graph cuts. IEEE Transactions on Pattern Analysis and Machine Intel- ligence, 23:1222–1239.

Descombes, X. (2011).Stochastic geometry for image anal- ysis. Wiley/Iste, x. descombes edition.

Descombes, X., Minlos, R., and Zhizhina, E. (2009). Object extraction using a stochastic birth-and-death dynamics in continuum. Journal of Mathematical Imaging and Vision, 33(3):347–359.

Dong, G. and Acton, S. (2007). Detection of rolling leuko- cytes by marked point processes. Journal of Elec- tronic Imaging, 16:033013.

Gamal-Eldin, A., Descombes, X., Charpiat, G., and Zeru- bia, J. (2011). A fast multiple birth and cut algorithm using belief propagation. InImage Processing (ICIP), 2011 18th IEEE International Conference on, pages 2813–2816. IEEE.

Gamal Eldin, A., Descombes, X., Charpiat, G., Zerubia, J., et al. (2012). Multiple birth and cut algorithm for mul- tiple object detection.Journal of Multimedia Process- ing and Technologies.

Green, P. J. (1995). Reversible jump markov chain monte carlo computation and bayesian model determination.

Biometrika, 82:711–732.

Kolmogorov, V. and Zabih, R. (2004). What energy func- tions can be minimized via graph cuts. IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 26:65–81.

Nesterov, Y. (2004). Introductory lectures on convex opti- mization: A basic course, volume 87. Springer.

Perrin, G., Descombes, X., and Zerubia, J. (2005). A marked point process model for tree crown extraction in plantations. InImage Processing, 2005. ICIP 2005.

IEEE International Conference on, volume 1, pages I–661. IEEE.